Merge pull request #1958 from vbystricky:ocl_filter2D

modules/imgproc/src/filter.cpp

@@ -41,6 +41,7 @@
 //M*/
 
 #include "precomp.hpp"
+#include "opencl_kernels.hpp"
 
 /****************************************************************************************\
                                     Base Image Filter
@@ -3115,6 +3116,206 @@ template<typename ST, class CastOp, class VecOp> struct Filter2D : public BaseFi
 
 }
 
+namespace cv
+{
+
+#define DIVUP(total, grain) (((total) + (grain) - 1) / (grain))
+#define ROUNDUP(sz, n)      ((sz) + (n) - 1 - (((sz) + (n) - 1) % (n)))
+
+// prepare kernel: transpose and make double rows (+align). Returns size of aligned row
+// Samples:
+//        a b c
+// Input: d e f
+//        g h i
+// Output, last two zeros is the alignment:
+// a d g a d g 0 0
+// b e h b e h 0 0
+// c f i c f i 0 0
+template <typename T>
+static int _prepareKernelFilter2D(std::vector<T>& data, const Mat &kernel)
+{
+    Mat _kernel; kernel.convertTo(_kernel, DataDepth<T>::value);
+    int size_y_aligned = ROUNDUP(kernel.rows * 2, 4);
+    data.clear(); data.resize(size_y_aligned * kernel.cols, 0);
+    for (int x = 0; x < kernel.cols; x++)
+    {
+        for (int y = 0; y < kernel.rows; y++)
+        {
+            data[x * size_y_aligned + y] = _kernel.at<T>(y, x);
+            data[x * size_y_aligned + y + kernel.rows] = _kernel.at<T>(y, x);
+        }
+    }
+    return size_y_aligned;
+}
+
+static bool ocl_filter2D( InputArray _src, OutputArray _dst, int ddepth,
+                   InputArray _kernel, Point anchor,
+                   double delta, int borderType )
+{
+    if (abs(delta) > FLT_MIN)
+        return false;
+
+    int type = _src.type();
+    int cn = CV_MAT_CN(type);
+    if ((1 != cn) && (2 != cn) && (4 != cn))
+        return false;//TODO
+
+    int sdepth = CV_MAT_DEPTH(type);
+    Size ksize = _kernel.size();
+    if( anchor.x < 0 )
+        anchor.x = ksize.width / 2;
+    if( anchor.y < 0 )
+        anchor.y = ksize.height / 2;
+    if( ddepth < 0 )
+        ddepth = sdepth;
+    else if (ddepth != sdepth)
+        return false;
+
+    bool isIsolatedBorder = (borderType & BORDER_ISOLATED) != 0;
+    bool useDouble = (CV_64F == sdepth);
+    const cv::ocl::Device &device = cv::ocl::Device::getDefault();
+    int doubleFPConfig = device.doubleFPConfig();
+    if (useDouble && (0 == doubleFPConfig))
+        return false;
+
+    const char* btype = NULL;
+    switch (borderType & ~BORDER_ISOLATED)
+    {
+    case BORDER_CONSTANT:
+        btype = "BORDER_CONSTANT";
+        break;
+    case BORDER_REPLICATE:
+        btype = "BORDER_REPLICATE";
+        break;
+    case BORDER_REFLECT:
+        btype = "BORDER_REFLECT";
+        break;
+    case BORDER_WRAP:
+        return false;
+    case BORDER_REFLECT101:
+        btype = "BORDER_REFLECT_101";
+        break;
+    }
+
+    cv::Mat kernelMat = _kernel.getMat();
+    std::vector<float> kernelMatDataFloat;
+    std::vector<double> kernelMatDataDouble;
+    int kernel_size_y2_aligned = useDouble ?
+            _prepareKernelFilter2D<double>(kernelMatDataDouble, kernelMat)
+            : _prepareKernelFilter2D<float>(kernelMatDataFloat, kernelMat);
+
+
+    cv::Size sz = _src.size();
+    size_t globalsize[2] = {sz.width, sz.height};
+    size_t localsize[2] = {0, 1};
+
+    ocl::Kernel kernel;
+    UMat src; Size wholeSize;
+    if (!isIsolatedBorder)
+    {
+        src = _src.getUMat();
+        Point ofs;
+        src.locateROI(wholeSize, ofs);
+    }
+
+    size_t maxWorkItemSizes[32]; device.maxWorkItemSizes(maxWorkItemSizes);
+    size_t tryWorkItems = maxWorkItemSizes[0];
+    for (;;)
+    {
+        size_t BLOCK_SIZE = tryWorkItems;
+        while (BLOCK_SIZE > 32 && BLOCK_SIZE >= (size_t)ksize.width * 2 && BLOCK_SIZE > (size_t)sz.width * 2)
+            BLOCK_SIZE /= 2;
+#if 1 // TODO Mode with several blocks requires a much more VGPRs, so this optimization is not actual for the current devices
+        size_t BLOCK_SIZE_Y = 1;
+#else
+        size_t BLOCK_SIZE_Y = 8; // TODO Check heuristic value on devices
+        while (BLOCK_SIZE_Y < BLOCK_SIZE / 8 && BLOCK_SIZE_Y * src.clCxt->getDeviceInfo().maxComputeUnits * 32 < (size_t)src.rows)
+            BLOCK_SIZE_Y *= 2;
+#endif
+
+        if ((size_t)ksize.width > BLOCK_SIZE)
+            return false;
+
+        int requiredTop = anchor.y;
+        int requiredLeft = (int)BLOCK_SIZE; // not this: anchor.x;
+        int requiredBottom = ksize.height - 1 - anchor.y;
+        int requiredRight = (int)BLOCK_SIZE; // not this: ksize.width - 1 - anchor.x;
+        int h = isIsolatedBorder ? sz.height : wholeSize.height;
+        int w = isIsolatedBorder ? sz.width : wholeSize.width;
+        bool extra_extrapolation = h < requiredTop || h < requiredBottom || w < requiredLeft || w < requiredRight;
+
+        if ((w < ksize.width) || (h < ksize.height))
+            return false;
+
+        char build_options[1024];
+        sprintf(build_options, "-D LOCAL_SIZE=%d -D BLOCK_SIZE_Y=%d -D DATA_DEPTH=%d -D DATA_CHAN=%d -D USE_DOUBLE=%d "
+                "-D ANCHOR_X=%d -D ANCHOR_Y=%d -D KERNEL_SIZE_X=%d -D KERNEL_SIZE_Y=%d -D KERNEL_SIZE_Y2_ALIGNED=%d "
+                "-D %s -D %s -D %s",
+                (int)BLOCK_SIZE, (int)BLOCK_SIZE_Y,
+                sdepth, cn, useDouble ? 1 : 0,
+                anchor.x, anchor.y, ksize.width, ksize.height, kernel_size_y2_aligned,
+                btype,
+                extra_extrapolation ? "EXTRA_EXTRAPOLATION" : "NO_EXTRA_EXTRAPOLATION",
+                isIsolatedBorder ? "BORDER_ISOLATED" : "NO_BORDER_ISOLATED");
+
+        localsize[0] = BLOCK_SIZE;
+        globalsize[0] = DIVUP(sz.width, BLOCK_SIZE - (ksize.width - 1)) * BLOCK_SIZE;
+        globalsize[1] = DIVUP(sz.height, BLOCK_SIZE_Y);
+
+        cv::String errmsg;
+        if (!kernel.create("filter2D", cv::ocl::imgproc::filter2D_oclsrc, build_options))
+            return false;
+        size_t kernelWorkGroupSize = kernel.workGroupSize();
+        if (localsize[0] <= kernelWorkGroupSize)
+            break;
+        if (BLOCK_SIZE < kernelWorkGroupSize)
+            return false;
+        tryWorkItems = kernelWorkGroupSize;
+    }
+
+    _dst.create(sz, CV_MAKETYPE(ddepth, cn));
+    UMat dst = _dst.getUMat();
+    if (src.empty())
+        src = _src.getUMat();
+
+    int idxArg = 0;
+    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(src));
+    idxArg = kernel.set(idxArg, (int)src.step);
+
+    int srcOffsetX = (int)((src.offset % src.step) / src.elemSize());
+    int srcOffsetY = (int)(src.offset / src.step);
+    int srcEndX = (isIsolatedBorder ? (srcOffsetX + sz.width) : wholeSize.width);
+    int srcEndY = (isIsolatedBorder ? (srcOffsetY + sz.height) : wholeSize.height);
+    idxArg = kernel.set(idxArg, srcOffsetX);
+    idxArg = kernel.set(idxArg, srcOffsetY);
+    idxArg = kernel.set(idxArg, srcEndX);
+    idxArg = kernel.set(idxArg, srcEndY);
+
+    idxArg = kernel.set(idxArg, ocl::KernelArg::WriteOnly(dst));
+    float borderValue[4] = {0, 0, 0, 0};
+    double borderValueDouble[4] = {0, 0, 0, 0};
+    if ((borderType & ~BORDER_ISOLATED) == BORDER_CONSTANT)
+    {
+        int cnocl = (3 == cn) ? 4 : cn;
+        if (useDouble)
+            idxArg = kernel.set(idxArg, (void *)&borderValueDouble[0], sizeof(double) * cnocl);
+        else
+            idxArg = kernel.set(idxArg, (void *)&borderValue[0], sizeof(float) * cnocl);
+    }
+    if (useDouble)
+    {
+        UMat kernalDataUMat(kernelMatDataDouble, true);
+        idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(kernalDataUMat));
+    }
+    else
+    {
+        UMat kernalDataUMat(kernelMatDataFloat, true);
+        idxArg = kernel.set(idxArg, ocl::KernelArg::PtrReadOnly(kernalDataUMat));
+    }
+    return kernel.run(2, globalsize, localsize, true);
+}
+}
+
 cv::Ptr<cv::BaseFilter> cv::getLinearFilter(int srcType, int dstType,
                                 InputArray filter_kernel, Point anchor,
                                 double delta, int bits)
@@ -3230,6 +3431,10 @@ void cv::filter2D( InputArray _src, OutputArray _dst, int ddepth,
                    InputArray _kernel, Point anchor,
                    double delta, int borderType )
 {
+    bool use_opencl = ocl::useOpenCL() && _dst.isUMat();
+    if( use_opencl && ocl_filter2D(_src, _dst, ddepth, _kernel, anchor, delta, borderType))
+        return;
+
     Mat src = _src.getMat(), kernel = _kernel.getMat();
 
     if( ddepth < 0 )

modules/imgproc/src/opencl/filter2D.cl

@@ -0,0 +1,375 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2010-2013, Advanced Micro Devices, Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors as is and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#ifdef BORDER_REPLICATE
+//BORDER_REPLICATE:     aaaaaa|abcdefgh|hhhhhhh
+#define ADDR_L(i, l_edge, r_edge)  ((i) <  (l_edge) ? (l_edge)   : (i))
+#define ADDR_R(i, r_edge, addr)    ((i) >= (r_edge) ? (r_edge)-1 : (addr))
+#define ADDR_H(i, t_edge, b_edge)  ((i) <  (t_edge) ? (t_edge)   :(i))
+#define ADDR_B(i, b_edge, addr)    ((i) >= (b_edge) ? (b_edge)-1 :(addr))
+#endif
+
+#ifdef BORDER_REFLECT
+//BORDER_REFLECT:       fedcba|abcdefgh|hgfedcb
+#define ADDR_L(i, l_edge, r_edge)  ((i) <  (l_edge) ? -(i)-1               : (i))
+#define ADDR_R(i, r_edge, addr)    ((i) >= (r_edge) ? -(i)-1+((r_edge)<<1) : (addr))
+#define ADDR_H(i, t_edge, b_edge)  ((i) <  (t_edge) ? -(i)-1 : (i))
+#define ADDR_B(i, b_edge, addr)    ((i) >= (b_edge) ? -(i)-1+((b_edge)<<1) : (addr))
+#endif
+
+#ifdef BORDER_REFLECT_101
+//BORDER_REFLECT_101:   gfedcb|abcdefgh|gfedcba
+#define ADDR_L(i, l_edge, r_edge)  ((i) <  (l_edge) ? -(i)                 : (i))
+#define ADDR_R(i, r_edge, addr)    ((i) >= (r_edge) ? -(i)-2+((r_edge)<<1) : (addr))
+#define ADDR_H(i, t_edge, b_edge)  ((i) <  (t_edge) ? -(i)                 : (i))
+#define ADDR_B(i, b_edge, addr)    ((i) >= (b_edge) ? -(i)-2+((b_edge)<<1) : (addr))
+#endif
+
+//blur function does not support BORDER_WRAP
+#ifdef BORDER_WRAP
+//BORDER_WRAP:          cdefgh|abcdefgh|abcdefg
+#define ADDR_L(i, l_edge, r_edge)  ((i) <  (l_edge) ? (i)+(r_edge) : (i))
+#define ADDR_R(i, r_edge, addr)    ((i) >= (r_edge) ? (i)-(r_edge) : (addr))
+#define ADDR_H(i, t_edge, b_edge)  ((i) <  (t_edge) ? (i)+(b_edge) : (i))
+#define ADDR_B(i, b_edge, addr)    ((i) >= (b_edge) ? (i)-(b_edge) : (addr))
+#endif
+
+#ifdef EXTRA_EXTRAPOLATION // border > src image size
+#ifdef BORDER_CONSTANT
+// None
+#elif defined BORDER_REPLICATE
+#define EXTRAPOLATE(x, y, minX, minY, maxX, maxY) \
+    { \
+        x = max(min(x, maxX - 1), minX); \
+        y = max(min(y, maxY - 1), minY); \
+    }
+#elif defined BORDER_WRAP
+#define EXTRAPOLATE(x, y, minX, minY, maxX, maxY) \
+    { \
+        if (x < minX) \
+            x -= ((x - maxX + 1) / maxX) * maxX; \
+        if (x >= maxX) \
+            x %= maxX; \
+        if (y < minY) \
+            y -= ((y - maxY + 1) / maxY) * maxY; \
+        if (y >= maxY) \
+            y %= maxY; \
+    }
+#elif defined(BORDER_REFLECT) || defined(BORDER_REFLECT_101)
+#define EXTRAPOLATE_(x, y, minX, minY, maxX, maxY, delta) \
+    { \
+        if (maxX - minX == 1) \
+            x = minX; \
+        else \
+            do \
+            { \
+                if (x < minX) \
+                    x = minX - (x - minX) - 1 + delta; \
+                else \
+                    x = maxX - 1 - (x - maxX) - delta; \
+            } \
+            while (x >= maxX || x < minX); \
+        \
+        if (maxY - minY == 1) \
+            y = minY; \
+        else \
+            do \
+            { \
+                if (y < minY) \
+                    y = minY - (y - minY) - 1 + delta; \
+                else \
+                    y = maxY - 1 - (y - maxY) - delta; \
+            } \
+            while (y >= maxY || y < minY); \
+    }
+#ifdef BORDER_REFLECT
+#define EXTRAPOLATE(x, y, minX, minY, maxX, maxY) EXTRAPOLATE_(x, y, minX, minY, maxX, maxY, 0)
+#elif defined(BORDER_REFLECT_101)
+#define EXTRAPOLATE(x, y, minX, minY, maxX, maxY) EXTRAPOLATE_(x, y, minX, minY, maxX, maxY, 1)
+#endif
+#else
+#error No extrapolation method
+#endif
+#else
+#define EXTRAPOLATE(x, y, minX, minY, maxX, maxY) \
+    { \
+        int _row = y - minY, _col = x - minX; \
+        _row = ADDR_H(_row, 0, maxY - minY); \
+        _row = ADDR_B(_row, maxY - minY, _row); \
+        y = _row + minY; \
+        \
+        _col = ADDR_L(_col, 0, maxX - minX); \
+        _col = ADDR_R(_col, maxX - minX, _col); \
+        x = _col + minX; \
+    }
+#endif
+
+#if USE_DOUBLE
+#ifdef cl_amd_fp64
+#pragma OPENCL EXTENSION cl_amd_fp64:enable
+#elif defined (cl_khr_fp64)
+#pragma OPENCL EXTENSION cl_khr_fp64:enable
+#endif
+#define FPTYPE double
+#define CONVERT_TO_FPTYPE CAT(convert_double, VEC_SIZE)
+#else
+#define FPTYPE float
+#define CONVERT_TO_FPTYPE CAT(convert_float, VEC_SIZE)
+#endif
+
+#if DATA_DEPTH == 0
+#define BASE_TYPE uchar
+#elif DATA_DEPTH == 1
+#define BASE_TYPE char
+#elif DATA_DEPTH == 2
+#define BASE_TYPE ushort
+#elif DATA_DEPTH == 3
+#define BASE_TYPE short
+#elif DATA_DEPTH == 4
+#define BASE_TYPE int
+#elif DATA_DEPTH == 5
+#define BASE_TYPE float
+#elif DATA_DEPTH == 6
+#define BASE_TYPE double
+#else
+#error data_depth
+#endif
+
+#define __CAT(x, y) x##y
+#define CAT(x, y) __CAT(x, y)
+
+#define uchar1 uchar
+#define char1 char
+#define ushort1 ushort
+#define short1 short
+#define int1 int
+#define float1 float
+#define double1 double
+
+#define convert_uchar1_sat_rte convert_uchar_sat_rte
+#define convert_char1_sat_rte convert_char_sat_rte
+#define convert_ushort1_sat_rte convert_ushort_sat_rte
+#define convert_short1_sat_rte convert_short_sat_rte
+#define convert_int1_sat_rte convert_int_sat_rte
+#define convert_float1
+#define convert_double1
+
+#if DATA_DEPTH == 5 || DATA_DEPTH == 6
+#define CONVERT_TO_TYPE CAT(CAT(convert_, BASE_TYPE), VEC_SIZE)
+#else
+#define CONVERT_TO_TYPE CAT(CAT(CAT(convert_, BASE_TYPE), VEC_SIZE), _sat_rte)
+#endif
+
+#define VEC_SIZE DATA_CHAN
+
+#define VEC_TYPE CAT(BASE_TYPE, VEC_SIZE)
+#define TYPE VEC_TYPE
+
+#define SCALAR_TYPE CAT(FPTYPE, VEC_SIZE)
+
+#define INTERMEDIATE_TYPE CAT(FPTYPE, VEC_SIZE)
+
+struct RectCoords
+{
+    int x1, y1, x2, y2;
+};
+
+//#define DEBUG
+#ifdef DEBUG
+#define DEBUG_ONLY(x) x
+#define ASSERT(condition) do { if (!(condition)) { printf("BUG in boxFilter kernel (global=%d,%d): " #condition "\n", get_global_id(0), get_global_id(1)); } } while (0)
+#else
+#define DEBUG_ONLY(x) (void)0
+#define ASSERT(condition) (void)0
+#endif
+
+
+inline INTERMEDIATE_TYPE readSrcPixel(int2 pos, __global const uchar* srcptr, int srcstep, const struct RectCoords srcCoords
+#ifdef BORDER_CONSTANT
+               , SCALAR_TYPE borderValue
+#endif
+    )
+{
+#ifdef BORDER_ISOLATED
+    if(pos.x >= srcCoords.x1 && pos.y >= srcCoords.y1 && pos.x < srcCoords.x2 && pos.y < srcCoords.y2)
+#else
+    if(pos.x >= 0 && pos.y >= 0 && pos.x < srcCoords.x2 && pos.y < srcCoords.y2)
+#endif
+    {
+        //__global TYPE* ptr = (__global TYPE*)((__global char*)src + pos.x * sizeof(TYPE) + pos.y * srcStepBytes);
+        __global TYPE* ptr = (__global TYPE*)(srcptr + pos.y * srcstep + pos.x * sizeof(TYPE));
+        return CONVERT_TO_FPTYPE(*ptr);
+    }
+    else
+    {
+#ifdef BORDER_CONSTANT
+        return borderValue;
+#else
+        int selected_col = pos.x;
+        int selected_row = pos.y;
+
+        EXTRAPOLATE(selected_col, selected_row,
+#ifdef BORDER_ISOLATED
+                srcCoords.x1, srcCoords.y1,
+#else
+                0, 0,
+#endif
+                srcCoords.x2, srcCoords.y2
+         );
+
+        // debug border mapping
+        //printf("pos=%d,%d --> %d, %d\n", pos.x, pos.y, selected_col, selected_row);
+
+        pos = (int2)(selected_col, selected_row);
+        if(pos.x >= 0 && pos.y >= 0 && pos.x < srcCoords.x2 && pos.y < srcCoords.y2)
+        {
+            //__global TYPE* ptr = (__global TYPE*)((__global char*)src + pos.x * sizeof(TYPE) + pos.y * srcStepBytes);
+            __global TYPE* ptr = (__global TYPE*)(srcptr + pos.y * srcstep + pos.x * sizeof(TYPE));
+            return CONVERT_TO_FPTYPE(*ptr);
+        }
+        else
+        {
+            // for debug only
+            DEBUG_ONLY(printf("BUG in boxFilter kernel\n"));
+            return (FPTYPE)(0.0f);
+        }
+#endif
+    }
+}
+
+// INPUT PARAMETER: BLOCK_SIZE_Y (via defines)
+
+__kernel
+__attribute__((reqd_work_group_size(LOCAL_SIZE, 1, 1)))
+void filter2D(__global const uchar* srcptr, int srcstep, int srcOffsetX, int srcOffsetY, int srcEndX, int srcEndY,
+                __global uchar* dstptr, int dststep, int dstoffset,
+               int rows, int cols,
+#ifdef BORDER_CONSTANT
+               SCALAR_TYPE borderValue,
+#endif
+               __constant FPTYPE* kernelData // transposed: [KERNEL_SIZE_X][KERNEL_SIZE_Y2_ALIGNED]
+               )
+{
+    const struct RectCoords srcCoords = {srcOffsetX, srcOffsetY, srcEndX, srcEndY}; // for non-isolated border: offsetX, offsetY, wholeX, wholeY
+
+    const int local_id = get_local_id(0);
+    const int x = local_id + (LOCAL_SIZE - (KERNEL_SIZE_X - 1)) * get_group_id(0) - ANCHOR_X;
+    const int y = get_global_id(1) * BLOCK_SIZE_Y;
+
+    INTERMEDIATE_TYPE data[KERNEL_SIZE_Y];
+    __local INTERMEDIATE_TYPE sumOfCols[LOCAL_SIZE];
+
+    int2 srcPos = (int2)(srcCoords.x1 + x, srcCoords.y1 + y - ANCHOR_Y);
+
+    int2 pos = (int2)(x, y);
+    __global TYPE* dstPtr = (__global TYPE*)((__global char*)dstptr + pos.y * dststep + dstoffset + pos.x * sizeof(TYPE)); // Pointer can be out of bounds!
+    bool writeResult = ((local_id >= ANCHOR_X) && (local_id < LOCAL_SIZE - (KERNEL_SIZE_X - 1 - ANCHOR_X)) &&
+                        (pos.x >= 0) && (pos.x < cols));
+
+#if BLOCK_SIZE_Y > 1
+    bool readAllpixels = true;
+    int sy_index = 0; // current index in data[] array
+
+    dstRowsMax = min(rows, pos.y + BLOCK_SIZE_Y);
+    for (;
+         pos.y < dstRowsMax;
+         pos.y++,
+         dstPtr = (__global TYPE*)((__global char*)dstptr + dststep))
+#endif
+    {
+        ASSERT(pos.y < dstRowsMax);
+
+        for (
+#if BLOCK_SIZE_Y > 1
+            int sy = readAllpixels ? 0 : -1; sy < (readAllpixels ? KERNEL_SIZE_Y : 0);
+#else
+            int sy = 0, sy_index = 0; sy < KERNEL_SIZE_Y;
+#endif
+            sy++, srcPos.y++)
+        {
+            data[sy + sy_index] = readSrcPixel(srcPos, srcptr, srcstep, srcCoords
+#ifdef BORDER_CONSTANT
+                    , borderValue
+#endif
+                    );
+        }
+
+        INTERMEDIATE_TYPE total_sum = 0;
+        for (int sx = 0; sx < KERNEL_SIZE_X; sx++)
+        {
+            {
+                __constant FPTYPE* k = &kernelData[KERNEL_SIZE_Y2_ALIGNED * sx
+#if BLOCK_SIZE_Y > 1
+                                                   + KERNEL_SIZE_Y - sy_index
+#endif
+                                                   ];
+                INTERMEDIATE_TYPE tmp_sum = 0;
+                for (int sy = 0; sy < KERNEL_SIZE_Y; sy++)
+                {
+                    tmp_sum += data[sy] * k[sy];
+                }
+
+                sumOfCols[local_id] = tmp_sum;
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+
+            int id = local_id + sx - ANCHOR_X;
+            if (id >= 0 && id < LOCAL_SIZE)
+               total_sum += sumOfCols[id];
+
+            barrier(CLK_LOCAL_MEM_FENCE);
+        }
+
+        if (writeResult)
+        {
+            *dstPtr = CONVERT_TO_TYPE(total_sum);
+        }
+
+#if BLOCK_SIZE_Y > 1
+        readAllpixels = false;
+#if BLOCK_SIZE_Y > KERNEL_SIZE_Y
+        sy_index = (sy_index + 1 <= KERNEL_SIZE_Y) ? sy_index + 1 : 1;
+#else
+        sy_index++;
+#endif
+#endif // BLOCK_SIZE_Y == 1
+    }
+}

modules/imgproc/test/ocl/test_filter2d.cpp

@@ -0,0 +1,135 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved.
+// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved.
+// Copyright (C) 2010-2012, Multicoreware, Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#include "test_precomp.hpp"
+#include "opencv2/ts/ocl_test.hpp"
+
+#ifdef HAVE_OPENCL
+
+namespace cvtest {
+namespace ocl {
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+// Filter2D
+PARAM_TEST_CASE(Filter2D, MatDepth, Channels, BorderType, bool, bool)
+{
+    static const int kernelMinSize = 2;
+    static const int kernelMaxSize = 10;
+
+    int type;
+    Size dsize;
+    Point anchor;
+    int borderType;
+    bool useRoi;
+    Mat kernel;
+
+    TEST_DECLARE_INPUT_PARAMETER(src)
+    TEST_DECLARE_OUTPUT_PARAMETER(dst)
+
+    virtual void SetUp()
+    {
+        type = CV_MAKE_TYPE(GET_PARAM(0), GET_PARAM(1));
+        borderType = GET_PARAM(2) | (GET_PARAM(3) ? BORDER_ISOLATED : 0);
+        useRoi = GET_PARAM(4);
+    }
+
+    void random_roi()
+    {
+        dsize = randomSize(1, MAX_VALUE);
+
+        Size ksize = randomSize(kernelMinSize, kernelMaxSize);
+        Mat temp = randomMat(ksize, CV_MAKE_TYPE(((CV_64F == CV_MAT_DEPTH(type)) ? CV_64F : CV_32F), 1), -MAX_VALUE, MAX_VALUE);
+        cv::normalize(temp, kernel, 1.0, 0.0, NORM_L1);
+
+        Size roiSize = randomSize(ksize.width, MAX_VALUE, ksize.height, MAX_VALUE);
+        Border srcBorder = randomBorder(0, useRoi ? MAX_VALUE : 0);
+        randomSubMat(src, src_roi, roiSize, srcBorder, type, -MAX_VALUE, MAX_VALUE);
+
+        Border dstBorder = randomBorder(0, useRoi ? MAX_VALUE : 0);
+        randomSubMat(dst, dst_roi, dsize, dstBorder, type, -MAX_VALUE, MAX_VALUE);
+
+        anchor.x = randomInt(-1, ksize.width);
+        anchor.y = randomInt(-1, ksize.height);
+
+        UMAT_UPLOAD_INPUT_PARAMETER(src)
+        UMAT_UPLOAD_OUTPUT_PARAMETER(dst)
+    }
+
+    void Near(double threshold = 0.0)
+    {
+        EXPECT_MAT_NEAR(dst, udst, threshold);
+        EXPECT_MAT_NEAR(dst_roi, udst_roi, threshold);
+    }
+};
+
+OCL_TEST_P(Filter2D, Mat)
+{
+    for (int j = 0; j < test_loop_times; j++)
+    {
+        random_roi();
+
+        OCL_OFF(cv::filter2D(src_roi, dst_roi, -1, kernel, anchor, 0.0, borderType));
+        OCL_ON(cv::filter2D(usrc_roi, udst_roi, -1, kernel, anchor, 0.0, borderType));
+
+        Near(1.0);
+    }
+}
+
+
+OCL_INSTANTIATE_TEST_CASE_P(ImageProc, Filter2D,
+                            Combine(
+                                Values(CV_8U, CV_16U, CV_16S, CV_32F, CV_64F),
+                                Values(1, 2, 4),
+                                Values((BorderType)BORDER_CONSTANT,
+                                       (BorderType)BORDER_REPLICATE,
+                                       (BorderType)BORDER_REFLECT,
+                                       (BorderType)BORDER_REFLECT_101),
+                                Bool(), // BORDER_ISOLATED
+                                Bool()  // ROI
+                                )
+                           );
+
+
+} } // namespace cvtest::ocl
+
+#endif // HAVE_OPENCL